Amazon Wants the AI People in Your Listing Photos Tagged

If any image or A+ module on your Amazon listings shows a photorealistic person who does not exist, you now have two obligations, and only one of them has a published penalty. Amazon told third-party sellers in the week of July 20 that product images and A+ content containing photorealistic AI-generated people must carry metadata flagging them before upload, and that it will add a shopper-facing indicator to qualifying listings. Sitting underneath that is New York General Business Law section 396-b, in force since June 9, 2026, which requires a conspicuous disclosure in any commercial advertisement containing a "synthetic performer" and sets civil penalties of $1,000 for a first violation and $5,000 for each one after. Metadata answers Amazon. It is not obvious that metadata alone answers New York.

What New York actually wrote

The statute came from S8420-A, signed December 11, 2025 and effective 180 days later. It defines a synthetic performer as "a digitally created asset created, reproduced, or modified by computer, using generative artificial intelligence or a software algorithm, that is intended to create the impression that the asset is engaging in an audiovisual and/or visual performance of a human performer who is not recognizable as any identifiable natural performer."

Read that definition slowly, because three words in it do most of the work. "Human performer" means the rule is about depicted people, not AI-generated backgrounds, packaging renders or product shots. "Not recognizable as any identifiable natural performer" means it targets invented faces rather than edited real ones. And the disclosure duty attaches where the business has actual knowledge that a synthetic performer is in the advertisement, which is a low bar for a seller who commissioned the image.

The exemptions are narrow and mostly will not help a merchant: advertisements for expressive works such as film, television, streaming and video games where the use is consistent with the work itself; audio-only advertisements; and cases where AI is used solely to translate the language of a human performer. Distribution media such as newspapers and networks are exempt unless they receive written notice and fail to stop distribution within five days.

The part that catches sellers outside New York is reach. As the law firm Crowell and Moring put it in its client alert, the obligation runs to "any company whose ads reach consumers in New York, regardless of where the advertiser is headquartered." A listing page is visible in every state. If you sell nationally, you are in scope. Crowell also notes it is watching for guidance from New York authorities, which is a polite way of saying nobody has yet published what a compliant disclosure looks like on a product detail page.

What Amazon has published, and what it has not

The reported requirement is narrow and mechanical: apply metadata to product images and to A+ content, the enhanced graphics and video modules on listing pages, before uploading them. Forbes describes it as embedding an IPTC-compatible metadata keyword flagging synthetic performers in qualifying assets. The carve-outs, per the syndicated CNBC report, cover content depicting television, movie or video game characters, and real people whose likenesses were modified with AI tools.

Here is the honest split between what you can act on and what you cannot:

Take the literal keyword from the notification in your own Seller Central account. Several outlets have printed a string; we could not open a source that showed it in Amazon's own words, so we are not repeating it here. This is the ordinary shape of a platform rule arriving before its documentation, and it is the same dynamic we wrote about when Shopify gave vape merchants 14 days: the policy is enforceable long before the help page is complete.

Five questions that decide whether this touches you

Work through these in order for each asset. A "no" at any step ends the analysis for that file.

1. Does the asset depict a human? AI-generated rooms, textures, shadows and packaging renders are outside the statutory definition, which is written around a performer.
2. Is that human photorealistic? Illustration and obvious stylisation do not create the impression of a real person.
3. Was the person generated rather than edited? A real model whose photo you retouched with AI tools is excluded from Amazon's requirement as reported, and falls outside the "not recognizable as any identifiable natural performer" language.
4. Is the asset an advertisement rather than an expressive work? Product images and A+ modules are advertising. The film and games carve-out is not available to you.
5. Can a shopper in New York see it? On Amazon, yes.

Writing the tag so it survives your export

The underlying standard is worth learning, because it outlives whatever Amazon calls its field. The IPTC, which maintains the photo metadata standard the whole industry uses, published guidance for synthetic media in May 2023. The field is Digital Source Type, and the value for imagery made by trained AI models is trainedAlgorithmicMedia, identified by the URI http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. It travels in an XMP packet inside the image or video file, or in a C2PA manifest. There is a separate value, compositeSynthetic, for assets that blend generated elements with camera capture, which is exactly what a composited lifestyle shot is.

Two practical warnings. First, metadata is fragile: many editors and resizing pipelines strip XMP on export, so write the field last, then reopen the exported file and confirm the value is still there before it goes anywhere near a bulk upload. Second, a file-level tag is invisible to a shopper. New York asks for a conspicuous disclosure in the advertisement. Until New York authorities or Amazon say otherwise, the conservative reading is that the metadata is Amazon's requirement and a visible line, in an image or in the A+ copy, is what addresses the statute.

What non-compliance could cost, arithmetically

This calculation is ours, and it is an upper bound rather than a forecast. The statute prices a first violation at $1,000 and each subsequent violation at $5,000. A seller with 12 listings carrying invented models would, if each listing counted separately, face $1,000 plus 11 times $5,000, which is $56,000. Nothing we could open defines whether a violation is counted per listing, per asset, per campaign or per enforcement action, so treat the number as what it is: a demonstration that the penalty structure escalates fast enough that a dozen legacy listings is not a rounding error. Against that, re-tagging an existing catalogue is metadata work measured in an afternoon.

The wider point is one we made when AI imagery got cheap enough for every store to use: the constraint on AI product images was never the generation cost. It was accuracy, returns, and now disclosure. And the speed at which a new obligation arrived by email is a reminder to run the platform dependency audit on any channel that can change your compliance burden without asking you.

Before your next upload

1. Open the Seller Central notification and copy the exact keyword Amazon specifies. Do not work from a news article.
2. Filter your catalogue for assets showing people, then apply the five questions above. Most sellers will find the affected set is small and concentrated in lifestyle images.
3. Write the IPTC Digital Source Type value plus Amazon's keyword into those files, export, and verify the metadata survived.
4. Decide your position on a visible disclosure and apply it consistently across images and A+ copy rather than case by case.
5. Add one line to your product photography brief: any supplier delivering generated people must deliver the metadata with them.